combinatorial geometry of point sets with collinearitiesdavidwo/files/payne-phd.pdfpin strictly...
TRANSCRIPT
-
Combinatorial geometry of
point sets with collinearities
Michael S. Payne
Department of Mathematics and Statistics
The University of Melbourne
Submitted in total fulfilment of the requirements
of the degree of Doctor of Philosophy
February 2014
-
Abstract
In this thesis we study various combinatorial problems relating to the geom-
etry of point sets in the Euclidean plane. The unifying theme is that all the
problems involve point sets that are not in general position, but have some
collinearities. As well as giving rise to natural and interesting problems, the
study of point sets with collinearities has important connections to other
areas of mathematics such as number theory.
Dirac conjectured that every set P of n non-collinear points in the plane
contains a point in at least n2 − c lines determined by P , for some constantc. It is known that some point is in Ω(n) lines determined by P . We show
that some point is in at least n37 lines determined by P .
Erdős posed the problem to determine the maximum integer f(n, `) such
that every set of n points in the plane with at most ` collinear contains a sub-
set of f(n, `) points with no three collinear. First we prove that if ` 6 O(√n)
then f(n, `) > Ω(√n/ ln `). Second we prove that if ` 6 O(n(1−�)/2) then
f(n, `) > Ω(√n log` n), which implies all previously known lower bounds on
f(n, `) and improves them when ` is not constant. Our results answer a sym-
metric version of the problem posed by Gowers, namely how many points
are required to ensure there are q collinear points or q points in general
position.
The visibility graph of a finite set of points in the plane has an edge between
two points if the line segment between them contains no other points. We
establish bounds on the edge- and vertex-connectivity of visibility graphs.
We find that every minimum edge cut is the set of edges incident to a vertex
of minimum degree. For vertex-connectivity, we prove that every visibility
graph with n vertices and at most ` collinear vertices has connectivity at
least n−1`−1 , which is tight. We also prove that the vertex-connectivity is at
least half the minimum degree.
We study some questions related to bichromatic point sets in the plane.
Given two disjoint point sets A and B in the plane, the bivisibility graph
1
-
has an edge between a point in A and a point in B if there are no other
points on the line segment between them. We characterise the connected
components of bivisibility graphs and give lower bounds on the number of
edges and the maximum degree. We also show that all sufficiently large
visibility graphs contain a given bipartite graph or many collinear points.
Lastly we make some progress on a conjecture of Kleitman and Pinchasi
about lower bounds on the number of bichromatic lines determined by a
bichromatic point set.
An empty pentagon in a point set P in the plane is a set of five points in
P in strictly convex position with no other point of P in their convex hull.
We prove that every finite set of at least 328`2 points in the plane contains
an empty pentagon or ` collinear points. This bound is optimal up to a
constant factor.
2
-
Declaration
This is to certify that:
• this thesis comprises only my original work towards the PhD exceptwhere indicated in the Preface,
• due acknowledgement has been made in the text to all other materialused, and
• this thesis is fewer than one hundred thousand words in length, exclu-sive of tables, maps, bibliographies and appendices.
Michael S. Payne
3
-
Preface
All work towards this thesis was carried out during the period of PhD candi-
dature at the University of Melbourne. None of the work has been submitted
for any other qualification. Except for results of other authors who are ac-
knowledged as they are introduced, the results of Chapters 3 through 7 are
to the best of my knowledge original contributions.
Most of this work is the result of academic collaboration which I gratefully
acknowledge. Much of this work has been published in a peer reviewed
journal or is publicly available as a preprint and currently under peer re-
view. In each case I was primarily responsible for the planning, drafting and
preparation of the work for publication.
Chapter 3 is the result of collaboration with my thesis advisor David Wood.
It consists mostly of material from our preprint Progress on Dirac’s con-
jecture [72] which is currently under peer review. Sections 3.2.1 and 3.2.2
contain additional material.
Chapter 4 is the result of collaboration with David Wood. It consists en-
tirely of material from our paper On the general position subset selection
problem [71], with some revisions.
Chapter 5 is the result of collaboration with Attila Pór, Pavel Valtr and
David Wood. It consists entirely of material from our paper On the connec-
tivity of visibility graphs [70], with some revisions.
The material of Chapter 6 is my own work and has not been published
elsewhere. Parts of it can be considered as extensions to Chapters 3 and 5.
Chapter 7 is the result of collaboration with János Barát, Vida Dujmović,
Gwenaël Joret, Ludmila Scharf, Daria Schymura, Pavel Valtr and David
Wood. It consists entirely of material from our preprint Empty pentagons in
point sets with collinearities [6] which is currently under peer review, with
some revisions.
All figures were created by myself except for Figures 2.1, 5.1, 7.8 and 7.9
4
-
which were created by David Wood, and the remaining figures of Chapter 7
which were created by Ludmila Scharf and Daria Schymura.
5
-
Acknowledgements
I would like to thank my advisor David Wood for his time and encourage-
ment, and all my co-authors with whom I enjoyed doing this work so much.
Thanks to the Group of Eight (Go8) and the German Academic Exchange
Service (DAAD) for funding our project Problems in geometric graph theory
which led in particular to the work in Chapter 7. Thanks to Jens Schmidt
and Helmut Alt for organising the German side of the project, and for
hosting me on various occasions at the Freie Universität Berlin. I am also
grateful to the Australian Government for providing for my living expenses
in the form of an Australian Postgraduate Award.
Thanks to my partner Anuradhi for her love and support and for sharing
the life of the graduate student with me – it’s been a lot of fun. Thanks
also to Moritz and Kaie, and all my friends from Berlin, for their hospitality
and company during my visits. And lastly thanks to my family for their
support, and especially my parents for encouraging my curiosity at every
stage.
6
-
Contents
1 Summary 11
2 Background 17
2.1 Incidence geometry in the plane 19
2.2 Independent sets in hypergraphs 25
2.3 Visibility graphs 26
2.4 Convex configurations 30
3 Dirac’s Conjecture and Beck’s Theorem 32
3.1 Dirac’s Conjecture 32
3.2 Beck’s Theorem 37
4 General position subset selection 43
4.1 Original problem 43
4.2 Generalised problem 47
4.3 Conjectures 50
5 Connectivity of visibility graphs 53
7
-
5.1 Edge connectivity 55
5.2 A key lemma 61
5.3 Vertex connectivity 64
5.4 Vertex connectivity with bounded collinearities 67
6 Bivisibility graphs 74
6.1 Connectedness of bivisibility graphs 74
6.2 Number of edges and complete bipartite subgraphs 75
6.3 Kleitman–Pinchasi Conjecture 78
7 Empty pentagons 85
7.1 Large subsets in weakly convex position 87
7.2 The empty edge lemma 94
7.3 Proof of Theorem 7.1 96
8
-
List of Figures
2.1 The graphs G2, G3, G4, G5 in the case of the 5× 5 grid. 22
5.1 A visibility graph with vertex-connectivity 2δ+13 . The black
vertices are a cut set. The minimum degree δ = 3k + 1 is
achieved, for example, at the top left vertex. Not all edges
are drawn. 54
5.2 If each ray from v through V (G) contains ` vertices, the de-
gree of v is n−1`−1 . 58
5.3 In each case the remaining points of B ∪ C must lie on thesolid segments of the rays. 60
5.4 Two properly coloured non-crossing geometric graphs with no
black-white edge between them. 61
5.5 Proof of Lemma 5.7. The shaded areas are empty. (a) A
type-1 visible pair. (b) A type-2 visible pair. (c) The highest
visible pair. (d) The lowest pair is type-1. (e) The lowest pair
is type-2. 64
5.6 Covering A ∪ B with rays and segments (a), each of whichcontains an edge of the bivisibility graph (b). 69
5.7 The only case in which h may not be perturbed to separate
the points assigned above h from those assigned below. 71
9
-
5.8 (a) The elliptic curve y2 = x3 − x. (b) The black pointsseparate the white points from the grey points. 73
6.1 Construction for Proposition 6.22. 83
7.1 The shaded regions represent the 4-sector S(p1, p2, p3, p4),
which may be bounded or unbounded. 87
7.2 (a) If |A ∩ b+| 6 |B ∩ l(b)|, then A is not minimal. (b) If b+
contained three non-collinear points of A, there would be an
empty pentagon. 88
7.3 Lemma 7.6. 90
7.4 Definition of b1 and the quadrilaterals Qi. 91
7.5 (a) If em is good then B ⊆ em. (b) If ep−1 is good and ep isnot, then B ⊆ ep−1. 92
7.6 (a) If vh ∈ bh then B ⊆ eh . (b) If vh+1 ∈ bh then A ∩⋃hi=1 b
+i = {v1, . . . , vh}. 93
7.7 The convex hull of B is covered by the union of the closed
sectors Si. 94
7.8 (a) Double-aligned. (b) Left-aligned. (c) Right-aligned. 95
7.9 (a) Neither double-aligned nor left-aligned nor right-aligned.
(b) The empty pentagon xj−2yj−2yj−1yjxj−1. 96
7.10 (a) The right child q and the left child p of v. (b) The quadri-
lateral Q(vq) and the sector S[vq]. 98
7.11 (a) u ∈ c⊕m and u ∈ d⊕n . (b) u cannot be in both d−i and c−i+1. 100
10
-
Chapter 1
Summary
In this thesis, various problems in combinatorial geometry are studied. The
unifying theme that runs through the problems is that they deal with finite
point sets in the Euclidean plane that have some collinearities. This chapter
contains a brief outline of the main contributions of the thesis. Many def-
initions, along with detailed discussion of the background to this work will
be deferred until Chapter 2.
Dirac’s Conjecture and Beck’s Theorem
Chapters 3 and 4 deal with combinatorial problems about incidences be-
tween points and lines in the Euclidean plane.
Dirac [17] conjectured that every set P of n non-collinear points in the
plane contains a point in at least n2 − c1 lines determined by P , for someconstant c1. The following weakening was proved by Beck [7] and Szemerédi–
Trotter [93]: every set P of n non-collinear points contains a point in at leastnc2
lines determined by P , for some large positive constant c2. In Chapter 3.1
we find new bounds on the constant c2. Our main result is the following.
Theorem (3.3). Every set P of n non-collinear points in the plane contains
a point in at least n37 lines determined by P .
11
-
In the same paper, Beck [7] proved that every set of n points in the plane
with at most ` collinear determines at least 1c3n(n− `) lines, for some largepositive constant c3. This is one of a pair of related results that are now
known as Beck’s Theorem. In Chapter 3.2 we calculate the best known
constant for Beck’s Theorem, proving the following theorem.
Theorem (3.15). Every set P of n points with at most ` collinear determines
at least 193n(n− `) lines.
General position subset selection
In Chapter 4.1 we study the problem of selecting a set in general position
from a set of points with collinearities, as originally posed by Erdős [28, 29].
Let f(n, `) be the maximum integer such that every set of n points in the
plane with at most ` collinear contains a subset of f(n, `) points with no
three collinear. We prove two main theorems.
Theorem (4.3). Let P be a set of n points with at most ` collinear. Then P
contains a set of Ω(n/√n ln `+ `2) points in general position. In particular,
if ` 6 O(√n) then P contains a set of Ω(
√nln `) points in general position.
Theorem (4.5). Fix constants � > 0 and d > 0. Let P be a set of n points
in the plane with at most ` collinear points, where ` 6 (dn)(1−�)/2. Then P
contains a set of Ω(√n log` n) points in general position.
Theorem 4.5 implies all previously known lower bounds on f(n, `) and im-
proves them when ` is not constant. Theorem 4.3 provides an almost com-
plete answer to a symmetric Ramsey style version of the general position
subset selection problem posed by Gowers [40]. He asked for the minimum
integer GP(q) such that every set of at least GP(q) points in the plane con-
tains q collinear points or q points in general position. Gowers noted that
GP(q) > Ω(q2), and Theorem 4.3 implies that GP(q) 6 O(q2 ln q), so the
asymptotic growth is determined up to a logarithmic factor.
In Chapter 4.2 we consider the more general problem of finding subsets with
at most k collinear points in a point set with at most ` collinear, and prove
12
-
analogous results in this setting too. Let f(n, `, k) be the maximum integer
such that every set of n points in the plane with at most ` collinear contains
a subset of f(n, `, k) points with at most k collinear.
Theorem (4.7). If k > 3 is constant and ` 6 O(√n), then
f(n, `, k) > Ω
(n(k−1)/k
`(k−2)/k
).
Theorem (4.9). Fix constants d > 0 and � ∈ (0, 1). If k > 3 is constantand 4 6 ` 6 dn(1−�)/2 then
f(n, `, k) > Ω
(n(k−1)/k
`(k−2)/k(lnn)1/k
).
There is a natural generalisation of Gowers’ problem to finding subsets with
at most k collinear. Let GPk(q) be the minimum integer such that every
set of at least GPk(q) points in the plane contains q collinear points or q
points with at most k collinear, for k > 3. It is not too hard to show that
GPk(q) > Ω(q2), and Theorem 4.7 implies that GPk(q) 6 O(q2), so in this
case the asymptotic growth is determined up to a constant factor.
Connectivity of visibility graphs
In studying point sets and the lines they generate, it is often useful to
consider the visibility graph of the point set. The visibility graph of a finite
set of points in the plane has the points as vertices and an edge between two
vertices if the line segment between them contains no other points in the
set. In Chapter 5 we study visibility graphs in their own right, focussing on
edge- and vertex-connectivity.
Unless all its vertices are collinear, a visibility graph has diameter at most 2,
and so it follows by a result of Plesńık [75] that its edge-connectivity equals
its minimum degree. We strengthen the result of Plesńık as follows.
Theorem (5.2). Let G be a graph with diameter 2. Then the edge-connec-
tivity of G equals its minimum degree. Moreover, for all distinct vertices
13
-
v and w in G, if d := min{deg(v), deg(w)} then there are d edge-disjointvw-paths of length at most 4.
Furthermore, we characterise minimum edge-cuts in visibility graphs.
Theorem (5.6). Every minimum edge-cut in a non-collinear visibility graph
is the set of edges incident to some vertex.
For vertex-connectivity, we prove the following.
Theorem (5.11). Every non-collinear visibility graph with minimum degree
δ has vertex-connectivity at least δ2 + 1.
Then we consider once again the parameter `, the maximum number of
collinear points.
Theorem (Corollary 5.15). Let G be the visibility graph of a set of n points
with at most ` collinear. Then G has vertex-connectivity at least n−1`−1 , which
is best possible.
In the case that ` = 4, we improve the bound in Theorem 5.11.
Theorem (5.18). Let G be a visibility graph with minimum degree δ and at
most four collinear vertices. Then G has vertex-connectivity at least 2δ+13 .
Theorem 5.18 is best possible for every δ since there are point sets with
at most three collinear points whose visibility graph has connectivity 2δ+13 .
The construction is due to Alperin, Buhler, Chalcraft and Rosenberg (see
Trimble [96] for an account of the authorship of the construction). It uses
real points on an elliptic curve and takes advantage of the group structure
that exists on these points. It is described at the end of Chapter 5.
Bivisibility graphs
In Chapter 6 we study a kind of bipartite visibility graph that was useful in
the investigation of the connectivity of visibility graphs. Given two disjoint
14
-
point sets A and B in the plane, the bivisibility graph has vertices A∪B andan edge between a point in A and a point in B if there are no other points
of A ∪B on the line segment between them.
The number of edges in a bivisibility graph is at least the number of lines
with a point from both A and B. These are called bichromatic lines. We
apply Theorem 3.15 and some other known results [65, 80] to obtain the
following lower bound on the number of bichromatic lines.
Theorem (Corollary 6.10). Let P be a set of n red and n blue points in
the plane with at most ` collinear. Then P determines at least 1186n(2n− `)bichromatic lines.
This also gives a lower bound on the maximum degree of a bivisibility graph.
Corollary (6.11). Let A be a set of n red points and B a set of n blue points
in the plane, such that A ∪ B is not collinear. Then the bivisibility graphB(A,B) has maximum degree at least n/94.
Another corollary is related to an important conjecture of Kára, Pór and
Wood [48]. The Big-Line-Big-Clique Conjecture asserts, roughly, that every
sufficiently large visibility graph contains a large clique or many collinear
points. Applying a classical result of Kővári, Sós and Turán [53] yields the
following bipartite subgraph version. A similar statement holds for bivisi-
bility graphs.
Corollary (6.14). For all integers t, ` > 2, there exists an integer n such
that every visibility graph on n or more points contains a Kt,t subgraph or `
collinear points.
We also make some progress toward a conjecture of Kleitman and Pin-
chasi [50].
Theorem (6.20 and 6.23). Let P be a set of n red, and n or n − 1 bluepoints in the plane. If neither colour class is collinear, then P determines
at least |P | − 2 bichromatic lines. Moreover, if n > 10, then P determinesat least |P | − 1 bichromatic lines, which is best possible.
15
-
Kleitman and Pinchasi conjectured that, under these assumptions, P deter-
mines at least |P | − 1 bichromatic lines for all n.
Empty pentagons
In Chapter 7 we study special configurations of points within point sets with
collinearities. We focus on empty convex k-gons, which are sets of k points in
strictly convex position with no other point in the convex hull. It is known
that a point set P , even in general position, need not contain an empty
heptagon no matter how large P is [45]. On the other hand, sufficiently
large point sets in general position always contain empty hexagons [39, 64].
It is an open question whether the same holds for sufficiently large point
sets with no ` collinear points.
We study this question for the case of empty pentagons. Abel et al. [1]
showed that sufficiently large point sets with no ` collinear always contain
empty pentagons. Their bound on the necessary size of such a point set was
doubly exponential in `. We improve this bound as follows.
Theorem (7.1). Let P be a finite set of points in the plane. If P contains at
least 328`2 points, then P contains an empty pentagon or ` collinear points.
This is optimal up to a constant factor since the (`−1)×(`−1) grid containsno empty pentagon and no ` collinear points.
16
-
Chapter 2
Background
The topic of this thesis is the combinatorial geometry of finite sets of points
in the Euclidean plane with collinearities. The problems studied are combi-
natorial in that they involve estimating the number or size of certain discrete
substructures within these point sets.
Any two points in the plane determine a straight line, and a collinearity is
when three or more points lie on a line. In terms of the space of possible
coordinates for these sets of points, almost all (in the measure theoretic
sense) finite point sets contain no collinearities. Another way of saying
this is that a randomly chosen point set would have no collinearities with
probability 1. In this sense point sets with collinearities are special.
By virtue of having collinearities, the coordinates of the points satisfy cer-
tain algebraic relationships. In investigating combinatorial properties of
these point sets, the configurations that are extremal often seem to exhibit
strong symmetries. It is not surprising then that such problems often lead
to interesting links with algebra and number theory, though these links are
not the focus of this thesis.
The methods employed here are rather more combinatorial and geometric
in nature. Graph theory plays a prominent role, as do basic topological
methods, ideas from convex geometry, and more. This mix of techniques is
17
-
not uncommon in combinatorics, and many of these ideas will be introduced
only as needed.
But let us begin by formally defining some terms that recur throughout the
thesis. The Euclidean plane R2 is referred to simply as the plane. Let P bea finite set of points in the plane. In fact, P always denotes such a set. A
line that contains at least two points in P is said to be determined by P .
A set of points is collinear if it is contained in a line, otherwise it is non-
collinear. A set of points in the plane is in general position if it contains no
three collinear points. (Other notions of general position are possible, but
this is the only one we consider.) So, if P is in general position, then every
line determined by P contains exactly two points from P .
A graph, often denoted G = (V,E), consists of a finite set V called the ver-
tices of G together with a set E of two-element subsets of V called the edges
of G. Less formally, the vertices of a graph represent some objects, while the
edges represent some connection or relation between pairs of vertices. Two
vertices connected by an edge are said to be adjacent. Defined as above, a
graph has no loops (edges from a vertex to itself) or repeated edges. Graphs
are abstract objects, but when studying geometric problems we are often
interested in specific representations of them. One basic representation is a
drawing of a graph in the plane, with points representing the vertices and
arcs connecting two vertices representing edges. A graph is planar if it has
a drawing without any edges intersecting (except at shared vertices) and a
plane graph is a graph together with such a drawing. It is often convenient
to conflate a graph and its elements with their representations, as already
demonstrated in the last sentence, where edges were identified with their
representing arcs. For the sake of readability, such abuse is used whenever
it is unlikely to cause any confusion. Graph theory has a large amount
of associated terminology, too much to define here. Our usage follows the
standard text on the subject [16].
Graphs appear often in this thesis, usually arising from some geometric
situation. The following is a prime example. Let P be a finite set of points
in the plane. Two distinct points v and w in the plane are visible with
18
-
respect to P if no point in P lies on the open line segment vw. The visibility
graph1 of P has vertex set P , and two vertices are adjacent if and only
if they are visible with respect to P . In other words, the visibility graph
is obtained by considering the lines determined by P , and two points are
adjacent if they are consecutive on such a line.
2.1 Incidence geometry in the plane
Visibility graphs are both a useful tool and an interesting object of study in
their own right. Their most famous application is probably in Székely’s cel-
ebrated proof [91] of the Szemerédi-Trotter Theorem [93], though he did not
use the term visibility graph. His proof partly inspired our work on incidence
geometry presented in Chapter 3, and the Szemerédi–Trotter Theorem is a
key ingredient in the results of Chapter 4. Indeed the proof is so short that
we are able to explain it and its prerequisites completely in this section. We
begin with Euler’s Formula2, continue to the Crossing Lemma, and then
prove the Szemerédi–Trotter Theorem as well as an important related result
known as Beck’s Theorem.
One of the most basic results in discrete geometry is the invariance of the
Euler characteristic, which is really a special case of deep results in topology
(see for example [43]). Given a plane graph G, the regions of the compliment
of G are known as faces.
Theorem 2.1 (Euler’s Formula). For any connected plane graph G with n
vertices, m edges and f faces,
n−m+ f = 2 .
There are many different proofs of Theorem 2.1. They have been collected by
Eppstein [24], and one of the simplest is the following. It uses multigraphs,
1There are various other kinds of graph called visibility graphs, such as graphs defined
by visibility among vertices of a polygon. We consider only the kind defined here.2There are various results known as Euler’s Formula. More specifically, we refer to
Euler’s Polyhedral Formula, also known as the Euler characteristic.
19
-
meaning loops and multiple edges are allowed.
Proof. Proceed by induction on the number of edges. If there are no edges,
then G is a single vertex, and there is one face, so n − m + f = 2. Nowsuppose there is at least one edge e. If e joins two distinct vertices, contract
it to a single vertex. This reduces n and m by 1, while leaving f unchanged,
so n−m+f is unchanged. On the other hand, if e is a loop, it separates twofaces (by the Jordan curve theorem). Delete e and merge these two faces.
This reduces m and f by 1, while leaving n unchanged, so n−m+f is againunchanged.
The crossing number of a graph G, denoted by cr(G), is the minimum
number of crossings in a drawing of G. See [69, 92] for surveys on the
crossing number. The following lower bound on cr(G) was first proved by
Ajtai et al. [3] and Leighton [56] (with weaker constants). A simple proof
with better constants can be found in [2]. The following version is due to
Pach et al. [66].
Theorem 2.2 (Crossing Lemma). For every graph G with n vertices and
m > 10316 n edges,
cr(G) >1024m3
31827n2.
The following well known proof gives weaker constants using only Theo-
rem 2.1 [13].
Proof. Using the fact that each face has at least three edges, it follows
from Theorem 2.1 that a planar graph has at most 3n edges. Starting
with a drawing of G with the fewest possible crossings, and removing edges
until the graph is planar, it follows that cr(G) > m − 3n. Now considera randomly chosen3 induced subgraph H of G that includes each vertex
independently with probability p. The expected number of vertices in H
is pn, the expected number of edges is p2m, and the expected number of
3Since probabilistic arguments are not used in any of the new proofs in this thesis we
omit any proper introduction. See a standard text such as [5].
20
-
crossings (in the sub-drawing of the original drawing) is p4 cr(G). Hence by
linearity of expectation, p4 cr(G) > p2m− 3pn. Setting p = 4n/m, which isless than 1 if G has at least 4n edges, it follows that cr(G) > m3/64n2.
As already mentioned, Székely’s proof [91] of the Szemeredi–Trotter Theo-
rem [93] uses the Crossing Lemma and visibility graphs. In fact, in Chap-
ter 3 we employ a slight strengthening of the Szemerédi–Trotter Theorem,
and that is what we prove here. First we need a few more definitions. For
i > 2, an i-line is a line containing exactly i points in P . Let si be the num-
ber of i-lines. Let Gi be the spanning subgraph of the visibility graph of P
consisting of all edges in j-lines where j > i; see Figure 2.1 for an example.
Note that since each i-line contributes i− 1 edges, |E(Gi)| =∑
j>i(j− 1)sj .Part (a) of the following version of the Szemerédi–Trotter Theorem gives a
bound on |E(Gi)|, while part (b) is the well known version that bounds thenumber of j-lines for j > i.
Theorem 2.3 (Szemerédi–Trotter Theorem). Let α and β be positive con-
stants such that every graph H with n vertices and m > αn edges satisfies
cr(H) >m3
βn2.
Let P be a set of n points in the plane. Then
(a)∑j>i
(j − 1)sj 6 max{αn,
β n2
2(i− 1)2}
,
and (b)∑j>i
sj 6 max
{αn
i− 1 ,β n2
2(i− 1)3}
.
Proof. Suppose∑
j>i(j − 1)sj = |E(Gi)| > αn. Applying the version of theCrossing Lemma assumed in the statement of Theorem 2.3 to Gi,
cr(Gi) >|E(Gi)|3βn2
=(∑
j>i(j − 1)sj)2|E(Gi)|βn2
>(i− 1)2(∑j>i sj)2|E(Gi)|
βn2.
On the other hand, since two lines cross at most once,
cr(Gi) 6
(∑j>i sj2
)6
1
2
(∑j>i
sj
)2.
21
-
G = G2 G3
G4 G5
Figure 2.1: The graphs G2, G3, G4, G5 in the case of the 5× 5 grid.
Combining these inequalities yields part (a). Part (b) follows directly from
part (a).
In 1951, Dirac [17] conjectured that every set P of n non-collinear points
contains a point in at least n2 − c1 lines determined by P , for some constantc1. Ten years later Erdős [26] suggested a weakening of this conjecture,
that there must be a point in at least n/c2 lines determined by P , for
some constant c2 > 0. In the 1983 paper that established Theorem 2.3,
Szemerédi and Trotter [93] proved this weakening as a consequence of their
main theorem.
Independently and at the same time, Beck [7] also proved the weakened
conjecture using a result similar to the Szemerédi–Trotter Theorem but
22
-
somewhat weaker. Beck also used the following theorem, which is some-
times known as ‘Beck’s Theorem’. To distinguish it from the related result
Theorem 2.6 below, we will call it Beck’s Two Extremes Theorem4.
Theorem 2.4 (Beck’s Two Extremes Theorem). Let P be a set of n points
not all collinear. Then either (a) some line contains 2−15n points in P , or
(b) P determines at least 2−15n2 lines with at most 27 points.
Here we give a short and simple proof of Beck’s Two Extremes Theorem
using the Szemerédi–Trotter Theorem. It is based on a well known proof
(see for example [12]), with some refinements similar to those we use in
Chapter 3. No attempt is made to optimise the constants.
Proof. Note that if n < 216 we are done since alternative (a) must hold.
Suppose alternative (a) does not hold, so there are at most 2−15n points
on a line. Consider the number of pairs of points that determine lines with
more than 27 points. We will apply Theorem 2.3(a) with α = 4 and β = 64.
2−15n∑i=27+1
(i
2
)si =
1
2
2−15n∑i=27+1
i(i− 1)si
=1
2
27 2−15n∑i=27+1
(i− 1)si +2−15n∑j=27+1
2−15n∑i=j
(i− 1)si
6
1
2
βn22 · 27 + αn+
2−15n∑i=27
βn2
2i2+ αn
6 16n2
(2−7 +
∑i=27
1
i2
)+ 2−14n2
6 0.251n2 .
This implies that the number of pairs of points that determine lines with
at most 27 points is at least(n2
)− 0.251n2 = 0.249n2 − 0.5n > 0.24899n2,
since n > 216. Hence the number of such lines is at least 0.24899n2/(27
2
)>
2−15n2.4Here we follow Theran [95]
23
-
Erdős’ weakening of Dirac’s Conjecture is an immediate consequence. We
include Beck’s proof.
Theorem 2.5 (Weak Dirac Conjecture). Every set P of n non-collinear
points contains a point in at least 2−15n lines determined by P .
Proof. If alternative (b) of Theorem 2.4 holds, then some point is in at least
2−15n2/n lines. On the other hand, if alternative (a) holds, then there is a
line with at least 2−15n points, and any point not on this line is in at least
2−15n lines.
In Chapter 3.1 we make some progress towards Dirac’s Conjecture, showing
that 2−15 can be improved to 1/37 in the above theorem.
Apart from Theorem 2.5, Beck’s [7] other main result was the following
theorem, settling a conjecture of Erdős [27]. It is also a consequence of
Theorem 2.4. We include Beck’s proof for completeness.
Theorem 2.6 (Beck’s Theorem). Let P be a set of n points with at most `
collinear. Then P determines at least 2−31n(n− `) lines.
Proof. If alternative (b) of Theorem 2.4 holds then we are done. So suppose
that alternative (a) holds and the longest line L contains ` > 2−15n points
in P . There are n− ` points not on L, so let H be a set of h = 2−15(n− `)of them. Counting lines with one point in L and at least one in H (and
subtracting overcounts) yields the following lower bound on the number of
lines.
h`−(h
2
)> h
(`− h
2
)>n− `215
(n
215− n− `
216
)>n− `215
· n216
.
The constant in the above theorem is even weaker than what is typically
given. This is because our version of Theorem 2.4 was designed to improve
the constant in Theorem 2.5. However, in Chapter 3.2 we will see that the
24
-
constant in this version can easily be improved to 2−16. Much more careful
analysis allows us to improve the constant to 1/93.
To illustrate the broad importance of the Szemerédi–Trotter Theorem, we
pause to mention a notable application in number theory. Given a finite set
of real numbers A, the sum set A+A is the set of sums of pairs of numbers
in A, and the product set A · A is the set of products of pairs of numbersin A. It is a natural question how small these sets may be. When A is an
arithmetic progression, |A+ A| = 2|A| − 1, which is minimal. When A is ageometric progression, |A ·A| = 2|A|−1, which is minimal. However, as onemight suspect, it is not possible for both the sum set and the product set of
A to be small (linear in |A|). Erdős and Szemerédi [35] proved that, for somec, δ > 0, either the sum set or the product set has size at least c|A|1+δ. LaterElekes [21] improved this to 25 |A|5/4. His simple proof applied the Szemerédi–Trotter Theorem to a set of points and lines constructed from the set A.
Solymosi [84] improved the bound further through a more sophisticated
application of the Szemerédi–Trotter Theorem. Erdős and Szemerédi [35]
conjectured that, for all � > 0, either the sum set or the product set has size
at least c(�)|A|2−�, for some constant c(�) > 0.
2.2 Independent sets in hypergraphs
The Szemerédi–Trotter Theorem again turns out to be useful in Chapter 4
when we study the size of the largest general position subset in a point set
with collinearities. We use it to bound the number of collinear triples of
points in a point set as a function of its size and the maximum number
of collinear points. By focussing on collinear triples, we are able to apply
known results about uniform hypergraphs.
A hypergraph consists of a set of vertices along with a set of subsets of the
vertex set called hyperedges. A hypergraph is k-uniform if all the hyperedges
have cardinality k. Thus a graph is a 2-uniform hypergraph.
For a point set P in the plane, let H(P ) be the 3-uniform hypergraph having
25
-
vertex set P and a hyperedge for each collinear triple in P . Even if P has
more than three points on some line, H(P ) captures all the information
about collinearities in P . Collinear subsets of P are those subsets L such
that every triple of points in L is a hyperedge in H(P ).
An independent set in a hypergraph is a set of vertices that does not contain
any hyperedges. The independence number of a hypergraph H, denoted
α(H), is the size of the largest independent set in H. Since an independent
set in H(P ) corresponds to a subset of P with no three collinear points,
independent sets are precisely the subsets of P in general position. The
independence number of H(P ) is therefore the size of the largest subset of
P in general position.
Our general approach in Chapter 4 is to combine the bound on the number
of hyperedges obtained from the Szemerédi–Trotter Theorem with known
bounds on the independence number of 3-uniform hypergraphs. The bound
on the number of hyperedges encapsulates some of the geometric restrictions
on P , while the bounds on independence numbers are purely combinatorial.
They are generally proven using the probabilistic method. Since we use
these bounds as black boxes, we will only introduce them as needed.
2.3 Visibility graphs
We have seen some ways in which visibility graphs have played a role in the
development of discrete geometry. The study of visibility graphs in their own
right is a relatively recent development. Much of this study has focussed on
questions related to the clique and chromatic number of visibility graphs.
The chromatic number of a graph is the least number of colours required to
colour the vertices so that no two adjacent vertices receive the same colour.
A clique in a graph is a complete subgraph, that is, a set of vertices among
which every possible edge is present. The clique number of a graph is the
size of the largest clique. Clearly the chromatic number of a graph is at
least the clique number.
26
-
Kára, Pór and Wood [48] asked whether the chromatic number of a visibility
graph is bounded from above by a function of its clique number. They
showed that visibility graphs with clique number at most 3 are 3-colourable.
They also made the following important conjecture.
Conjecture 2.7 (Big-Line-Big-Clique Conjecture). For all integers k, ` >
2, there exists an integer n such that every visibility graph on n or more
vertices contains a clique of size k or ` collinear points.
So far this conjecture has only been proven for k 6 5 (see Section 2.4 below).
The most obvious approach for the general conjecture fails. Turán’s Theo-
rem [97] says that the maximum number of edges in a graph on n vertices
with no clique of size k is (k−1)n2
2k . However, for each n, Sylvester [87–90]
constructed a set of n points with no four collinear whose visibility graph hasn2
3 +O(n) edges. For large n this is less than the number of edges required
by Turán’s Theorem. See [48, 59, 73, 78] for more results and conjectures
about the clique and chromatic number of visibility graphs. Further related
results can be found in [1, 20].
Faced with the difficulty of proving Conjecture 2.7 and other related con-
jectures, we decided to study more basic properties of visibility graphs so
as to deepen our understanding of their structure. In Chapter 5 we inves-
tigate the connectivity properties of visibility graphs. A graph is connected
if there exists a path between any two vertices in the graph. A graph is
k-vertex-connected if it has more than k vertices and it remains connected
whenever fewer than k vertices are deleted. Since a complete graph cannot
be disconnected by removing vertices, this means that Kn is (n− 1)-vertex-connected. A graph is k-edge-connected if it remains connected whenever
fewer than k edges are deleted. A vertex (edge) cut in a graph is a set of ver-
tices (edges) whose removal disconnects the graph. Thus a (non-complete)
graph is k-connected if its cuts all have size at least k. Menger’s Theorem
gives a very useful characterisation of k-connectivity.
Theorem 2.8 (Menger’s Theorem). A graph is k-vertex-connected (k-edge-
connected) if and only if there exist k internally vertex-disjoint (edge-disjoint)
paths between each pair of distinct vertices.
27
-
The degree of a vertex v in a graph is the number of edges that contain v.
The minimum (maximum) degree of a graph G is the minimum (maximum)
degree of a vertex in G. Our main results in Chapter 5 give bounds on the
connectivity of visibility graphs in terms of the minimum degree.
A graph is bipartite if it has chromatic number at most 2. In other words,
the vertex set can be partitioned into two parts so that each edges contains
a vertex in each part. In Chapter 6 we study a kind of bipartite visibility
graph that was useful in Chapter 5. Given two disjoint point sets in the
plane A and B, the bivisibility graph has vertex set A ∪ B, and an edgebetween a vertex in A and another in B if they are visible with respect
to A ∪ B. The sets A and B are often thought of as being coloured withtwo different colours. We begin by characterising the connected components
of bivisibility graphs, and then turn our attention to lower bounds on the
number of edges. The number of edges in a bivisibility graph is at least the
number of bichromatic lines, that is, lines containing a point from both A
and B.
The study of bichromatic lines in bichromatic point sets has some history.
We use some results of Pach and Pinchasi [65] and Purdy and Smith [80] to
adapt our optimised version of Beck’s Theorem to give a lower bound on the
number of bichromatic lines. One corollary of this is a linear lower bound on
the maximum degree of non-collinear bivisibility graphs. Another corollary
is a bivisibility version of the Big-Line-Big-Clique Conjecture, which says
that sufficiently large bivisibility graphs with no ` collinear points contain
large complete bipartite subgraphs. Since bivisibility graphs are subgraphs
of visibility graphs, a similar statement holds for visibility graphs too. Unlike
the Big-Line-Big-Clique Conjecture, this corollary follows directly from well
known results in extremal graph theory.
Finally we turn our attention to general linear lower bounds on the number
of bichromatic lines, not depending on the maximum number of collinear
points. The monochromatic version of this problem has a longer history.
In 1948, de Bruijn and Erdős [15] proved that every non-collinear set of n
points in the plane determines at least n lines. In fact, they proved this
28
-
result in a more general combinatorial setting.
Theorem 2.9 (de Bruijn and Erdős). Let S be a set of cardinality n and
{S1, . . . , Sk} a collection of subsets of S such that each pair of elements inS is contained in exactly one Si. Then either S = Si for some i, or k > n.
As noted by de Bruijn and Erdős, the special case where S is a set of points
in the plane and the Si are the collinear subsets of S is easier to prove than
the general theorem. It follows by induction from the well-known Sylvester-
Gallai Theorem (actually first proven by Melchior [60]), which says that
every finite non-collinear set of points in the plane determines a line with
just two points. In the case of bichromatic lines, Kleitman and Pinchasi [50]
conjectured that if P is a set of n red, and n or n − 1 blue points in theplane and neither colour class is collinear, then P determines at least |P |−1bichromatic lines. As motivation, Kleitman and Pinchasi note that together
with the following theorem of Motzkin [63], their conjecture would imply
the plane case of Theorem 2.9.
Theorem 2.10 (Motzkin). Every non-collinear set of red and blue points
in the plane determines a monochromatic line.
We make some progress toward the conjecture of Kleitman and Pinchasi, but
also show that, unlike Theorem 2.9, it is not true in a purely combinatorial
setting. A similar combinatorial version of the problem has been studied by
Meshulam [61].
Theorem 2.11 (Meshulam). Let X1, . . . Xc be disjoint sets of cardinality n
(these are colour classes), let S =⋃iXi and let {S1, . . . , Sk} be a collection
of subsets of S such that each pair of elements in S is contained in exactly
one Si (these are ‘lines’). Then either S = Si for some i or |{i : ∀j Si 6⊂Xj}| > (c− 1)n (this counts non-monochromatic ‘lines’).
In the bichromatic case with c = 2 we have at least n bichromatic lines,
roughly half the number conjectured by Kleitman and Pinchasi under the
stronger assumption that no colour class is collinear. It is an interesting
question whether the lower bound of Theorem 2.11 can be improved under
this assumption.
29
-
2.4 Convex configurations
We now require some further geometric definitions. A set X in the plane is
convex if for every pair of points in X, the straight line segment between
them is also contained in X. For a set of points P in the plane, the convex
hull of P , denoted conv(P ), is the smallest convex set containing P . P is in
convex position if every point of P lies on the boundary of conv(P ). Another
classical result in discrete geometry is the Erdős–Szekeres Theorem [33].
Theorem 2.12 (Erdős–Szekeres Theorem). For every integer k there is
a minimum integer ES(k) such that every set of at least ES(k) points in
general position in the plane contains k points in convex position.
Erdős [27] asked whether a similar result held for empty k-gons, that is,
k points in convex position with no other points inside their convex hull.
Horton [45] answered this question in the negative by showing that there
are arbitrarily large point sets in general position that contain no empty
heptagon. On the other hand, Harborth [42] showed that every set of at least
10 points in general position contains an empty pentagon. More recently,
Nicolás [64] and Gerken [39] independently settled the question for k = 6 by
showing that sufficiently large point sets in general position always contain
empty hexagons; see also [52, 98].
In order to address similar questions for point sets with collinearities, it
is helpful to refine the definition of convex position. A point x ∈ P is acorner of P if conv(P \ {x}) 6= conv(P ). The set P is in strictly convexposition if every point in P is a corner of P . By way of contrast, a set in
convex position, but not necessarily in strictly convex position, is said to be
in weakly convex position. Thus a set in strictly convex position is also in
weakly convex position. A weakly (respectively strictly) convex k-gon is a
set of k points in weakly (respectively strictly) convex position.
It is well known that the Erdős–Szekeres theorem generalises for point sets
with collinearities; see [1] for proofs. One generalisation states that every set
of at least ES(k) points contains a weakly convex k-gon. For strictly convex
30
-
k-gons, it is necessary to consider point sets with bounded collinearities,
since a collinear point set has at most two points in strictly convex position.
In this case the generalisation states that for all integers k and ` there exists
a minimum integer ES(k, `) such that every set of at least ES(k, `) points in
the plane contains ` collinear points or a strictly convex k-gon.
In Chapter 7 we study the problem of finding strictly convex empty k-gons
in point sets with no ` collinear. Horton’s negative result [45] for empty
heptagons also applies in this setting. For k > 7 there may be no empty k-
gons even in a very large point set with bounded collinearities. On the other
hand, Abel et al. [1] showed that every finite set of at least ES((2`−1)`−1
2`−2
)points in the plane contains an empty pentagon or ` collinear points. The
case k = 6 remains open for ` > 4, and it is not clear how to adapt the proofs
of Nicolás [64] and Gerken [39] to deal with collinearities. Our contribution
is to improve on the result of Abel et al., showing that every finite set of at
least 328`2 points contains an empty pentagon or ` collinear points.
Note that since the vertices of an empty pentagon form a clique in the
visibility graph, this establishes the k 6 5 case of the Big-Line-Big-Clique
Conjecture (2.7). In the other direction, Wood [101] asked whether the vis-
ibility graphs of point sets with no empty pentagon have bounded clique or
chromatic number. Cibulka et al. [11] answered this question in the nega-
tive by constructing a family of sets with no empty pentagon but arbitrarily
large clique (and thus also chromatic) number.
31
-
Chapter 3
Dirac’s Conjecture and
Beck’s Theorem
3.1 Dirac’s Conjecture
In 1951, Gabriel Dirac [17] made the following conjecture, which remains
unresolved:
Conjecture 3.1 (Dirac’s Conjecture). There is a constant c1 such that
every set P of n non-collinear points contains a point in at least n2 − c1 linesdetermined by P .
See reference [4] for examples showing that the n2 bound would be tight.
Note that if P is non-collinear and contains at least n2 collinear points, then
Dirac’s Conjecture holds. Thus we may assume that P contains at mostn2 collinear points, and n > 5. In 1961, Erdős [26] proposed the following
weakened conjecture.
Conjecture 3.2 (Weak Dirac Conjecture). There is a constant c2 such that
every set P of n non-collinear points contains a point in at least nc2 lines
determined by P .
In 1983, the Weak Dirac Conjecture was proved independently by Beck [7]
32
-
and Szemerédi and Trotter [93], in both cases with c2 unspecified and very
large. We prove the Weak Dirac Conjecture with c2 much smaller. (See
references [30, 32, 49, 57, 79] for more on Dirac’s Conjecture.)
Theorem 3.3. Every set P of n non-collinear points contains a point in at
least n37 lines determined by P .
Theorem 3.3 is a consequence of the following theorem. The points of P
together with the lines determined by P are called the arrangement of P .
Theorem 3.4. For every set P of n points in the plane with at most n37
collinear points, the arrangement of P has at least n2
37 point-line incidences.
Proof of Theorem 3.3. Let P be a set of n non-collinear points in the plane.
If P contains at least n37 collinear points, then every other point is in at
least n37 lines determined by P (one through each of the collinear points).
Otherwise, by Theorem 3.4, the arrangement of P has at least n2
37 incidences,
and so some point is incident with at least n37 lines determined by P .
The proof of Theorem 3.4 takes inspiration from the well known proof of
Beck’s Two Extremes Theorem (2.4) [12] as a corollary of the Szemerédi–
Trotter Theorem (2.3) [93], and also from the simple proof of the Szemerédi–
Trotter Theorem due to Székely [91], which in turn is based on the Crossing
Lemma (2.2). These proofs were discussed in Chapter 2.
The proof of Theorem 3.4 also employs Hirzebruch’s Inequality [44]. As
before, si is the number of lines containing i points in P .
Theorem 3.5 (Hirzebruch’s Inequality). Let P be a set of n points with at
most n− 3 collinear. Then
s2 +3
4s3 > n+
∑i>5
(2i− 9)si .
Hirzebruch’s Inequality is rather interesting in that it does not follow from
Euler’s formula like many other results discussed here. Instead, it is a conse-
quence of deep results in algebraic geometry and it applies in a much broader
33
-
setting than the real plane. In particular, it is also valid for arrangements
of points in the complex plane. In 1995, Erdős and Purdy [31] asked for
a combinatorial proof of the inequality, a fascinating question that remains
open.
Theorem 3.4 follows from the Crossing Lemma (2.2) and the following gen-
eral result by setting α = 10316 , β =318271024 , c = 71, and δ = �, in which case
δ > 136.158 . The value of δ is readily calculated numerically since∑i>c
i+ 1
i3=∑i>1
i+ 1
i3−
c−1∑i=1
i+ 1
i3
= ζ(2) + ζ(3)−c−1∑i=1
i+ 1
i3
= 2.847 . . .−c−1∑i=1
i+ 1
i3,
where ζ is the Riemann zeta function.
Theorem 3.6. Let α and β be positive constants such that every graph H
with n vertices and m > αn edges satisfies
cr(H) >m3
βn2.
Fix an integer c > 8 and a real � ∈ (0, 12). Let h :=c(c−2)5c−18 . Then for every set
P of n points in the plane with at most �n collinear points, the arrangement
of P has at least δn2 point-line incidences, where
δ =1
h+ 1
(1− �α− β
2
((c− h− 2)(c+ 1)
c3+∑i>c
i+ 1
i3
)).
Proof. Let J := {2, 3, . . . , b�nc}. Considering the visibility graph G of P andits subgraphs Gi (as defined in Chapter 2), let k be the minimum integer
such that |E(Gk)| 6 αn. If there is no such k then let k := b�nc + 1. Aninteger i ∈ J is large if i > k, and is small if i 6 c. An integer in J that isneither large nor small is medium.
Recall that an i-line is a line containing i points in P . An i-pair is a pair
of points in an i-line. A small pair is an i-pair for some small i. Define
34
-
medium pairs and large pairs analogously, and let PS , PM and PL denote
the number of small, medium and large pairs respectively. An i-incidence
is an incidence between a point of P and an i-line. A small incidence is an
i-incidence for some small i. Define medium incidences analogously, and let
IS and IM denote the number of small and medium incidences respectively.
Let I denote the total number of incidences. Thus,
I =∑i∈J
isi .
The proof proceeds by establishing an upper bound on the number of small
pairs in terms of the number of small incidences. Analogous bounds are
proved for the number of medium pairs, and the number of large pairs.
Combining these results gives the desired lower bound on the total number
of incidences.
For the bound on small pairs, Hirzebruch’s Inequality (3.5) is useful. Since
we may assume fewer than n2 points are collinear, and thus n > 5, there
are no more than n− 3 collinear points. Therefore, Hirzebruch’s Inequalityimplies that hs2 +
3h4 s3 − hn− h
∑i>5(2i− 9)si > 0 since h > 0. Thus,
PS = s2 + 3s3 + 6s4 +c∑i=5
(i
2
)si
6 (h+ 1)s2 +
(3h
4+ 3
)s3 + 6s4 +
c∑i=5
(i
2
)si − hn− h
c∑i=5
(2i− 9)si
6h+ 1
2· 2s2 +
h+ 4
4· 3s3 +
3
2· 4s4 +
c∑i=5
(i− 1
2− 2h+ 9h
i
)isi − hn .
Setting X := max{h+12 ,
h+44 ,
32 ,max56i6c
(i−12 − 2h+ 9hi
)}implies that
PS 6 XIS − hn . (1)
The above inequality is strongest when X is minimised by determining the
optimal value of h as follows. Let γ(h, i) := i−12 − 2h + 9hi . The secondpartial derivative of γ(h, i) with respect to i is positive for i > 0, so γ(h, i)
is maximised for i = 5 or i = c, and the other values of i can be ignored.
Thus X is bounded from below by five linear functions of h. Notice that
35
-
for fixed c, h+12 increases with h, while γ(h, c) decreases with h. Therefore
X is at least the value of these functions at their intersection point, which
occurs at h = c(c−2)5c−18 . Using the fact that c > 8, it can be checked that this
intersection point satisfies the other three constraints1, and is therefore the
optimal solution.
To bound the number of medium pairs, consider a medium i ∈ J . Sincei is not large,
∑j>i(j − 1)sj > αn. Hence, using parts (a) and (b) of the
Szemerédi–Trotter Theorem (2.3),∑j>i
jsj =∑j>i
(j − 1)sj +∑j>i
sj 6βn2
2(i− 1)2 +βn2
2(i− 1)3 =βn2i
2(i− 1)3 . (2)
Given the factor X in the bound on the number of small pairs in (1), it
helps to introduce the same factor in the bound on the number of medium
pairs. It is convenient to define Y := c− 1− 2X.
PM −XIM =(
k−1∑i=c+1
(i
2
)si
)−X
(k−1∑i=c+1
isi
)
=1
2
k−1∑i=c+1
(i− 1− 2X) isi
=1
2
k−1∑i=c+1
(i− c+ Y ) isi
=1
2
k−1∑i=c+1
k−1∑j=i
jsj
+ Y2
(k−1∑i=c+1
isi
).
Applying (2) yields
PM −XIM 6β n2
4
(Yc+ 1
c3+∑i>c
i+ 1
i3
). (3)
It remains to bound the number of large pairs:
PL =
b�nc∑i=k
(i
2
)si 6
�n
2
∑i>k
(i− 1)si =�n
2|E(Gk)| 6
�α n2
2. (4)
1A simple way to do this is to note that h(c) increases with c for c > 8 and so h > 2411
.
Then compare X = h+12
to the other three constraints.
36
-
Combining (1), (3) and (4),(n
2
)=
1
2(n2 − n)
6 PS + PM + PL
6 XIS − hn+XIM +β n2
4
(Yc+ 1
c3+∑i>c
i+ 1
i3
)+�α n2
2.
Thus,
I > IS + IM >1
2X
(1− �α− β
2
(Yc+ 1
c3+∑i>c
i+ 1
i3
))n2 +
2h− 12X
n .
The result follows since h > 1.
It is worth noting that the methods used in the proof of Theorem 3.6 can
be used to obtain good lower bounds on the number of edges in a visibility
graph. The main difference is that edges (∑
(i − 1)si) are counted insteadof incidences (
∑isi). For instance, we can prove the following result.
Theorem 3.7. Let P be a set of n points in the plane with at most n50
collinear. Then the visibility graph of P has at least n2
50 edges.
For point sets with at most o(n) collinear points, the following is the best
asymptotic result we have obtained.
Theorem 3.8. Let P be a set of n points in the plane with at most `
collinear. Then the visibility graph of P has at least n2
39 −O(`n) edges.
3.2 Beck’s Theorem
In his work on the Weak Dirac Conjecture, Beck proved the following theo-
rem [7].
Theorem 3.9 (Beck’s Theorem). There is a constant c3 > 0 such that every
set P of n points with at most ` collinear determines at least c3n(n−`) lines.
37
-
In Chapter 2 we gave a relatively simple proof of Beck’s Theorem (2.6) with
c3 = 2−31. Here our aim is to find tighter bounds on c3. First we use
Theorem 3.6 and some well known lemmas to show that c3 > 198 . A more
tailored approach using similar methods is then employed to show c3 > 193 .
The first tool we need is a classical inequality due to Melchior [60]. The proof
uses Euler’s formula applied to the projective dual configuration. Melchior’s
Inequality was later rediscovered by Kelly and Moser [49].
Theorem 3.10 (Melchior’s Inequality). Let P be a set of n non-collinear
points. Then
s2 > 3 +∑i>4
(i− 3)si .
We will use the following straightforward corollary of Melchior’s Inequality.
As before, I is the total number of incidences in the arrangement of P . Let
E be the total number of edges in the visibility graph of P , and let L be the
total number of lines in the arrangement of P .
Lemma 3.11. If P is not collinear, then 3L > 3 + I, and 2L > 3 + E.
Proof. Melchior’s Inequality is often written∑
i>2(i−3)si 6 −3, which is tosay 3+
∑isi 6 3
∑si. Since I = E+L, it also follows that 2L > 3+E.
It is interesting to note that since I > 2L and E > L, all these parameters
are within a constant factor of each other.
When there is a large number of collinear points, the following lemma be-
comes useful.
Lemma 3.12. Let P be a set of n points in the plane such that some line
contains exactly ` points in P . Then the visibility graph of P contains at
least `(n− `) edges.
Proof. Let S be the set of ` collinear points in P . For each point v ∈ S andfor each point w ∈ P \ S, count the edge incident to w in the direction ofv. Since S is collinear and w is not in S, no edge is counted twice. Thus
E > |S| · |P \ S| = `(n− `).
38
-
We note in passing that Lemmas 3.11 and 3.12 can be used to improve
the proof of Theorem 2.6 and yield a constant of 2−16. However by using
Theorem 3.6 as well we can already do much better.
Theorem 3.13. Every set P of n points with at most ` collinear determines
at least 198n(n− `) lines.
Proof. Assume ` is the size of the largest collinear subset of P . If ` > n49 then
E > 149n(n − `) by Lemma 3.12 and thus L > 198n(n − `) by Lemma 3.11.On the other hand, suppose ` 6 n49 . Setting α =
10316 , β =
318271024 ,
�2 =
δ3
and c = 67 in Theorem 3.6 gives � > 149 and δ >1
32.57 . So I >1
32.57n2 >
132.57n(n− `) and thus L > 198n(n− `) by Lemma 3.11.
3.2.1 Further improvement
A more direct approach similar to the methods used in the proof of Theo-
rem 3.6 can be used to improve Theorem 3.13 slightly to yield 193n(n − `)lines. We use the following more general result, which again employs Hirze-
bruch’s Inequality (3.5).
Theorem 3.14. Let α and β be positive constants such that every graph H
with n vertices and m > αn edges satisfies
cr(H) >m3
βn2.
Fix an integer c > 29. Then for every set P of n points in the plane with at
most ` collinear points, the arrangement of P has at least(1
2− β
4
(1
c+∑i=c
1
i2
))4c− 16
c2 + 3c− 18n2 − α
2
4c− 16c2 + 3c− 18`n
lines with at most c points.
Proof. Define small, medium and large pairs and lines as in the proof of
Theorem 3.6. Then using Hirzebruch’s Inequality (3.5),
PS = s2 + 3s3 + 6s4 +c∑i=5
(i
2
)si
39
-
6 (h+ 1)s2 +
(3h
4+ 3
)s3 + 6s4 +
c∑i=5
(i
2
)si − hn− h
c∑i=5
(2i− 9)si
6 (h+ 1)s2 +3
4(h+ 4)s3 + 6s4 +
c∑i=5
(i(i− 1)
2− h(2i− 9)
)si − hn .
Using the fact that c > 29 and similar arguments to those used in the proof
of Theorem 3.6, it is advantageous to set h := c2−c−24c−16 . This gives
max
{h+ 1,
3
4(h+ 4), 6, max
56i6c
(i(i− 1)
2− h(2i− 9)
)}= h+ 1 =: X ,
and thus,
PS 6 XLS − hn .
For medium i, the assumed Crossing Lemma and part (a) of the Szemerédi-
Trotter Theorem (2.3), imply that∑j>i
(j − 1)sj 6βn2
2(i− 1)2 .
Thus,
PM =1
2
k∑i=c+1
i(i− 1)si
=1
2
c k∑i=c+1
(i− 1)si +k∑
j=c+1
k∑i=j
(i− 1)si
6
1
2
(βn2
2c+∑i=c+1
βn2
2(i− 1)2
)
6βn2
4
(1
c+∑i=c
1
i2
).
As in the proof of Theorem 3.6, we have PL 6 `αn/2. Adding it all up gives(n
2
)6 XLS +
βn2
4
(1
c+∑i=c
1
i2
)+`αn
2− hn ,
so (1
2− β
4
(1
c+∑i=c
1
i2
))n2 +
(h− 1
2− `α
2
)n 6 XLS ,
40
-
and since X = h+ 1 = c2+3c−184c−16 ,(
1
2− β
4
(1
c+∑i=c
1
i2
))4c− 16
c2 + 3c− 18n2 − `α
2
4c− 16c2 + 3c− 18n 6 LS .
For constant `, we may observe that the number of lines determined by P is
Ω(n2). Theorem 3.14 yields the best coefficient of n2. Setting c = 66 gives
at least 170n2 − 15`n lines with at most 66 points.
Lemma 3.12 together with Theorem 3.14 may be used to improve the con-
stant in Beck’s theorem further to 193 .
Theorem 3.15. Every set P of n points with at most ` collinear determines
at least 193n(n− `) lines.
Proof. We may assume that ` is the size of the longest line. If ` > �n for
some constant �, then by Lemmas 3.11 and 3.12, L > �n(n − `)/2. On theother hand, Theorem 3.14 says LS > An2 − Bn` for some A(c) and B(c)evident in the theorem. Observe that
2A
1 + 2B> �
=⇒ A > �/2 +B�− �2/2=⇒ An > �n/2 + (B − �/2)�n=⇒ An > �n/2 + (B − �/2)`=⇒ An2 −Bn` > �n(n− `)/2 .
So maximising 2A(c)1+2B(c) yields the best possible �. Setting c = 76 gives
� 6 1/46.2. Thus the constant for Beck’s Theorem is at least 1/92.4
3.2.2 Lines with few points
Beck’s Theorem is often stated as a bound on the number of lines with few
points. In his original paper, Beck [7] mentioned briefly in a footnote that
Lemma 3.11 implies the following.
41
-
Observation 3.16 (Beck). If P is not collinear, then at least half the lines
determined by P contain 3 points or less.
Proof. By Lemma 3.11,
3s2 + 3s3 + 3∑i>4
si >∑i>2
isi > 2s2 + 2s3 + 4∑i>4
si .
Thus
2(s2 + s3) >∑i>2
si ,
as desired.
Corollary 3.17. Every set P of n points with at most ` collinear determines
at least 1186n(n− `) lines each with at most 3 points.
Hirzebruch’s Inequality (3.5) may be used to find lower bounds on the num-
ber of lines with at most c points in a similar way to Observation 3.16.
Observation 3.18. The number of lines with at most c points for c > 4 is
at least 2c−72c−6 times the total number of lines.
Proof. If there are n − 2 collinear points then there is only one line withmore than c points and at least n− 1 lines with less than c points. We mayassume n > 5, so the lemma holds. If there are at most n−3 collinear pointsthen by Hirzebruch’s Inequality,
c∑i=2
si > s2 +3
4s3 >
∑i>5
(2i− 9)si > (2c− 7)∑i>c+1
si .
Thus,
(2c− 6)c∑i=2
si > (2c− 7)∑i>2
si .
42
-
Chapter 4
General position subset
selection
4.1 Original problem
Recall that a set of points in the plane is in general position if it contains no
three collinear points. The general position subset selection problem asks,
given a finite set of points in the plane with at most ` collinear, how big
is the largest subset in general position? That is, determine the maximum
integer f(n, `) such that every set of n points in the plane with at most `
collinear contains a subset of f(n, `) points in general position. Throughout
this chapter we assume ` > 3. Furthermore, as the results in this chapter
are all asymptotic in n, it will be made explicit whenever ` is a constant not
dependent on n. Otherwise ` is allowed to grow as a function of n.
The problem was originally posed by Erdős, first for the case ` = 3 [28],
and later in a more general form [29]. Füredi [36] showed that the density
version of the Hales–Jewett theorem [37] implies that f(n, `) 6 o(n) for all
`, and that a result of Phelps and Rödl [74] on independent sets in partial
Steiner triple systems implies that
f(n, 3) > Ω(√n lnn) .
43
-
Until recently, the best known lower bound for ` > 4 was√
2n/(`− 2),proved by a greedy selection algorithm. Lefmann [55] showed that for con-
stant `,
f(n, `) > Ω(√n lnn) .
(In fact, his results are more general, see Section 4.2.)
In relation to the general position subset selection problem (and its rel-
atives), Brass, Moser and Pach [9, p. 318] write, “To make any further
progress, one needs to explore the geometric structure of the problem.” We
do this by using the Szemerédi–Trotter Theorem (2.3).
We give improved lower bounds on f(n, `) when ` is not constant, with the
improvement being most significant for values of ` around√n. Our first
result (Theorem 4.3) says that if ` 6 O(√n) then f(n, `) > Ω(
√nln `). Our
second result (Theorem 4.5) says that if ` 6 O(n(1−�)/2) then f(n, `) >
Ω(√n log` n). For constant `, this implies Lefmann’s lower bound on f(n, `)
mentioned above.
Our main tool is the following lemma.
Lemma 4.1. Let P be a set of n points in the plane with at most ` collinear.
Then the number of collinear triples in P is at most c(n2 ln `+`2n) for some
constant c.
Proof. For 2 6 i 6 `, let si be the number of lines containing exactly i
points in P . The Szemerédi–Trotter Theorem (2.3) implies that for some
constant c > 1, and for all i > 2,∑j>i
sj 6 c
(n2
i3+n
i
).
Thus the number of collinear triples is
∑̀i=2
(i
3
)si 6
∑̀i=2
i2∑̀j=i
sj
6∑̀i=2
ci2(n2
i3+n
i
)
44
-
6 c∑̀i=2
(n2
i+ in
)6 c(n2 ln `+ `2n) .
Note that Lefmann [54] proved Lemma 4.1 for the case of the√n×√n grid
via a direct counting argument. A similar statement to Lemma 4.1 with
` =√n also appears in the book by Tao and Vu [94, Corollary 8.8].
To apply Lemma 4.1 it is useful to consider the 3-uniform hypergraph H(P )
determined by a set of points P , with vertex set P , and an edge for each
collinear triple in P . A subset of P is in general position if and only if it
is an independent set in H(P ). The size of the largest independent set in
a hypergraph H is denoted α(H). Spencer [85] proved the following lower
bound on α(H).
Lemma 4.2 (Spencer). Let H be an r-uniform hypergraph with n vertices
and m edges. If m < n/r then α(H) > n/2. If m > n/r then
α(H) >r − 1rr/(r−1)
n
(m/n)1/(r−1).
Lemmas 4.1 and 4.2 imply our first result.
Theorem 4.3. Let P be a set of n points with at most ` collinear. Then P
contains a set of Ω(n/√n ln `+ `2) points in general position. In particular,
if ` 6 O(√n) then P contains a set of Ω(
√nln `) points in general position.
Proof. Let m be the number of edges in H(P ). By Lemma 4.1, m/n 6
c(n ln ` + `2) for some constant c. Now apply Lemma 4.2 with r = 3. If
m < n/3 then α(H(P )) > n/2, as required. Otherwise,
α(H(P )) >2n
33/2(m/n)1/2>
2n
33/2√c(n ln `+ `2)
=2
3√
3c
n√n ln `+ `2
.
45
-
Note that Theorem 4.3 also shows that if `2/ ln ` > n then f(n, `) > Ω(n/`).
This improves upon the greedy bound mentioned in the introduction, and is
within a constant factor of optimal, since there are point sets with at most
` collinear that can be covered by n/` lines.
Theorem 4.3 answers, up to a logarithmic factor, a symmetric Ramsey style
version of the general position subset selection problem posed by Gowers [40].
He asked for the minimum integer GP(q) such that every set of at least
GP(q) points in the plane contains q collinear points or q points in general
position. Gowers noted that Ω(q2) 6 GP(q) 6 O(q3). Theorem 4.3 with
` = q − 1 and n = GP(q) implies that Ω(√
GP(q)/ ln(q − 1)) 6 q and soGP(q) 6 O(q2 ln q).
The bound GP(q) > Ω(q2) comes from the q × q grid, which contains noq + 1 collinear points, and no more than 2q + 1 in general position, since
each row can have at most 2 points. Determining the maximum number of
points in general position in the q × q grid is known as the no-three-in-lineproblem, first posed by Dudeney in 1917 [18]. See [41] for the best known
bound and for more on its history.
As an aside, note that Pach and Sharir [67] proved a result somewhat similar
to Lemma 4.1 for the number of triples in P determining a fixed angle
α ∈ (0, π). Their proof is similar to that of Lemma 4.1 in its use of theSzemerédi–Trotter theorem. Also, Elekes [22] employed Lemma 4.2 to prove
a similar result to Theorem 4.3 for the problem of finding large subsets
with no triple determining a given angle α ∈ (0, π). Pach and Sharir andElekes did not allow the case α = 0, that is, collinear triples. This may be
because their work did not consider the parameter `, without which the case
α = 0 is exceptional since P could be entirely collinear, and all triples could
determine the same angle.
The following lemma of Sudakov [86, Proposition 2.3] is a corollary of a
result by Duke, Lefmann and Rödl [19].
Lemma 4.4 (Sudakov). Let H be a 3-uniform hypergraph on n vertices with
m edges. Let t >√m/n and suppose there exists a constant � > 0 such that
46
-
the number of edges containing any fixed pair of vertices of H is at most
t1−�. Then α(H) > Ω(nt
√ln t)
.
Lemmas 4.1 and 4.4 can be used to prove our second result.
Theorem 4.5. Fix constants � > 0 and d > 0. Let P be a set of n points
in the plane with at most ` collinear points, where ` 6 (dn)(1−�)/2. Then P
contains a set of Ω(√n log` n) points in general position.
Proof. Let m be the number of edges in H(P ). By Lemma 4.1, for some
constant c > 1,
m 6 c`2n+ cn2 ln ` < cdn2 + cn2 ln ` 6 (d+ 1)cn2 ln ` .
Define t :=√
(d+ 1)cn ln `. Thus t >√m/n. Each pair of vertices in H is
in less than ` edges of H, and
` 6 (dn)(1−�)/2 < ((d+ 1)cn ln `)(1−�)/2 = t1−� .
Thus the assumptions in Lemma 4.4 are satisfied. So H contains an inde-
pendent set of size Ω(nt√
ln t). Moreover,
n
t
√ln t =
√n
(d+ 1)c ln `
√ln√
(d+ 1)cn ln `
>√
n
(d+ 1)c ln `
√1
2lnn
=
√1
2(d+ 1)c
√n lnn
ln `
= Ω(√n log` n) .
Thus P contains a subset of Ω(√n log` n) points in general position.
4.2 Generalised problem
In this section we consider a natural generalisation of the general position
subset selection problem. Given k < `, Erdős [29] asked for the maximum
47
-
integer f(n, `, k) such that every set of n points in the plane with at most `
collinear contains a subset of f(n, `, k) points with at most k collinear. Thus
f(n, `) = f(n, `, 2). We prove results similar to Theorems 4.3 and 4.5 in this
generalised setting.
Brass [8] considered this question for constant ` = k + 1, and showed that
o(n) > f(n, k + 1, k) > Ω(n(k−1)/k(lnn)1/k) .
This can be seen as a generalisation of the results of Füredi [36] for f(n, 3, 2).
As in Füredi’s work, the lower bound comes from a result on partial Steiner
systems [82], and the upper bound comes from the density Hales–Jewett
theorem [38]. Lefmann [55] further generalised these results for constant `
and k by showing that
f(n, `, k) > Ω(n(k−1)/k(lnn)1/k) .
The density Hales–Jewett theorem also implies the general bound f(n, `, k) 6
o(n) for all ` and k.
The result of Lefmann may be generalised to include the dependence of
f(n, `, k) on ` for constant k > 3, analogously to Theorems 4.3 and 4.5 for
k = 2. The first result we need is a generalisation of Lemma 4.1. It is proved
in the same way.
Lemma 4.6. Let P be a set of n points in the plane with at most ` collinear.
Then, for k > 4, the number of collinear k-tuples in P is at most c(`k−3n2 +
lk−1n) for some absolute constant c.
Lemmas 4.2 and 4.6 imply the following theorem which is proved in the
same way as Theorem 4.3.
Theorem 4.7. If k > 3 is constant and ` 6 O(√n), then
f(n, `, k) > Ω
(n(k−1)/k
`(k−2)/k
).
For ` =√n and constant k > 3, Theorem 4.7 implies
f(n,√n, k) > Ω
(n(k−1)/k
n(k−2)/2k
)= Ω
(n(2k−2−k+2)/2k
)= Ω(
√n) .
48
-
This answers completely a generalised version of Gowers’ question [40],
namely, to determine the minimum integer GPk(q) such that every set of at
least GPk(q) points in the plane contains q collinear points or q points with
at most k collinear, for k > 3. Thus GPk(q) 6 O(q2). The bound GPk(q) >
Ω(q2) comes from the following construction. Let m := b(q − 1)/kc and letP be the m×m grid. Then P has at most m points collinear, and m < q.If S is a subset of P with at most k collinear, then S has at most k points
in each row. So |S| 6 km 6 q − 1.
Theorem 4.5 can be generalised using Lemma 4.6 and a theorem of Duke,
Lefmann and Rödl [19] (the one that implies Lemma 4.4).
Theorem 4.8 (Duke, Lefmann and Rödl). Let H be a k-uniform hypergraph
with maximum degree ∆(H) 6 tk−1 where t� k. Let pj(H) be the numberof pairs of edges of H sharing exactly j vertices. If pj(H) 6 nt2k−j−1−γ for
j = 2, . . . , k − 1 and some constant γ > 0, then
α(H) > C(k, γ)n
t(lnn)1/(k−1)
for some constant C(k, γ) > 0.
Theorem 4.9. Fix constants d > 0 and � ∈ (0, 1). If k > 3 is constant and4 6 ` 6 dn(1−�)/2 then
f(n, `, k) > Ω
(n(k−1)/k
`(k−2)/k(lnn)1/k
).
Proof. Given a set P of n points with at most ` collinear, a subset with at
most k collinear points corresponds to an independent set in the (k + 1)-
uniform hypergraph Hk+1(P ) of collinear (k + 1)-tuples in P . By Lemma
4.6, the number of edges in Hk+1(P ) is m 6 c(n2`k−2 + nlk) for some
constant c. The first term dominates since ` 6 o(√n). For n large enough,
m/n 6 2cn`k−2.
To limit the maximum degree of Hk+1(P ), discard vertices of degree greater
than 2(k + 1)m/n. Let ñ be the number of such vertices. Considering
the sum of degrees, (k + 1)m > ñ2(k + 1)m/n, and so ñ 6 n/2. Thus
discarding these vertices yields a new point set P ′ such that |P ′| > n/2 and
49
-
∆(Hk+1(P′)) 6 4(k + 1)cn`k−2. Note that an independent set in Hk+1(P ′)
is also independent in Hk+1(P ).
Set t := (4(k + 1)cn`k−2)1/k, so m 6 12(k+1)ntk and ∆(Hk+1(P
′)) 6 tk, as
required for Theorem 4.8. By assumption, ` 6 dn(1−�)/2. Thus
` 6 d
(tk`2−k
4(k + 1)c
) 1−�2
.
Hence `2
1−�+k−2 6 d2/(1−�)tk
4(k+1)c , implying ` 6 C1(k)tk
21−�+k−2 = C1(k)t
1−�1−�+2�
k for
some constant C1(k). Define �′ := 1 − 1−�
1−�+ 2�k
, so �′ > 0 (since � < 1) and
` 6 C1(k)t1−�′. To bound pj(Hk+1(P
′)) for j = 2, . . . , k, first choose one
edge (which determines a line), then choose the subset to be shared, then
choose points from the line to complete the second edge of the pair. Thus
for γ := �′/2 and sufficiently large n,
pj(Hk+1(P′)) 6 m
(k + 1
j
)(`− k − 1k + 1− j
)6 C2(k)nt
k`k+1−j
6 C2(k)(C1(k))k+1−jntkt(1−�
′)(k+1−j)
6 nt2(k+1)−j−1−γ .
Hence the second requirement of Theorem 4.8 is satisfied. Thus
α(Hk+1(P′)) > Ω
(nt
(ln t)1/k)
> Ω
(n(k−1)/k
`(k−2)/k
(ln((n`k−2)1/k)
)1/k)
> Ω
(n(k−1)/k
`(k−2)/k(lnn)1/k
).
4.3 Conjectures
Theorem 4.7 suggests the following conjecture, which would completely an-
swer Gowers’ question [40], showing that GP(q) = Θ(q2). It is true for the√n×√n grid [25, 41].
50
-
Conjecture 4.10. f(n,√n) > Ω(
√n).
A natural variation of the general position subset selection problem is to
colour the points of P with as few colours as possible, such that each colour
class is in general position. A straightforward application of the Lovász
Local Lemma shows that under this requirement, n points with at most
` collinear are colourable with O(√`n) colours1. The following conjecture
would imply Conjecture 4.10. It is also true for the√n×√n grid [100].
Conjecture 4.11. Every set P of n points in the plane with at most√n
collinear can be coloured with O(√n) colours such that each colour class is
in general position.
The following proposition is somewhat weaker than Conjecture 4.11.
Proposition 4.12. Every set P of n points in the plane with at most√n
collinear can be coloured with O(√n ln3/2 n) colours such that each colour
class is in general position.
Proof. Colour P by iteratively selecting a largest subset in general position
and giving it a new colour. Let P0 := P . Let Ci be a largest subset of Pi
in general position and let Pi+1 := Pi \ Ci. Define ni := |Pi|. ApplyingLemma 4.1 to Pi shows that H(Pi) has O(n
2i ln ` + `
2ni) edges. Thus the
average degree of H(Pi) is at most O(ni ln ` + `2) which is O(n lnn) since
ni 6 n and ` 6√n.
Applying Lemma 4.2 gives |Ci| = α(H(Pi)) > cni/√n lnn for some constant
c > 0. Thus ni 6 n(1 − c/√n lnn)i. It is well known (and not difficult to
show) that if a sequence of numbers mi satisfies mi 6 m(1− 1/x)i for somex > 1 and if j > x lnm, then mj 6 1. Hence if k >
√n lnn lnn/c then
nk 6 1, so the number of colours used is O(√n ln3/2 n).
1Colouring each point of P with one of c colours uniformly at random, the probability
of a particular collinear triple being monochromatic is 1/c2. These events are independent
unless the triples intersect. Consider all lines determined by P that contain a fixed point
p, and let ki be the number of points on the ith line. Then the number of triples containing
p is∑i
(ki−1
2
)6 `
∑i ki 6 `n. Thus each triple intersects at most 3`n others. By the
Local Lemma there exists a proper colouring as long as 12`n 6 c2.
51
-
The problem of determining the correct asymptotics of f(n, `) (and f(n, `, k))
for constant ` remains wide open. The Szemerédi–Trotter theorem is essen-
tially tight for the√n ×√n grid [68], but says nothing for point sets with
bounded collinearities. For this reason, the lower bounds on f(n, `) for con-
stant ` remain essentially combinatorial. Finding a way to bring geometric
information to bear in this situation is an interesting challenge.
Conjecture 4.13. If ` is constant, then f(n, `) > Ω(n/polylog(n)).
The point set that gives the upper bound f(n, `) 6 o(n) (from the density
Hales–Jewett theorem) is the generic projection to the plane of the blog` nc-dimensional `×`×· · ·×` integer lattice (henceforth [`]d where d := blog`(n)c).The problem of finding large general position subsets in this point set for
` = 3 is known as Moser’s cube problem [62, 77], and the best known
asymptotic lower bound is Ω(n/√
lnn) [10, 77].
In the colouring setting, the following conjecture is equivalent to Conjec-
ture 4.13 by an argument similar to that of Proposition 4.12.
Conjecture 4.14. For constant ` > 3, every set of n points in the plane
with at most ` collinear can be coloured with O(polylog(n)) colours such that
each colour class is in general position.
Conjecture 4.14 is true for [`]d, which can be coloured with O(d`−1) colours
as follows. For each x ∈ [`]d, define a signature vector in Z` whose entries arethe number of entries in x equal to 1, 2, . . . `. The number of such signatures
is the number of partitions of d into at most ` parts, which is O(d`−1).
Give each set of points with the same signature its own colour. To see that
this is a proper colouring, suppose that {a, b, c} ⊂ [`]d is a monochromaticcollinear triple, with b between a and c. Permute the coordinates so that
the entries of b are non-decreasing. Consider the first coordinate i in which
ai, bi and ci are not all equal. Then without loss of generality, ai < bi. But
this implies that a has more entries equal to ai than b does, contradicting
the assumption that the signatures are equal.
52
-
Chapter 5
Connectivity of visibility
graphs
In this chapter we study the edge- and vertex-connectivity of visibility
graphs. A graph G on at least k + 1 vertices is k-vertex-connec